The Dynamic Kernel Scheduler—Part 1

Abstract

Emerging processor architectures such as GPUs and Intel MICs provide a huge performance potential for high performance computing. However developing software that uses these hardware accelerators introduces additional challenges for the developer. These challenges may include exposing increased parallelism, handling different hardware designs, and using multiple development frameworks in order to utilise devices from different vendors.The Dynamic Kernel Scheduler (DKS) is being developed in order to provide a software layer between the host application and different hardware accelerators. DKS handles the communication between the host and the device, schedules task execution, and provides a library of built-in algorithms. Algorithms available in the DKS library will be written in CUDA, OpenCL, and OpenMP. Depending on the available hardware, the DKS can select the appropriate implementation of the algorithm.The first DKS version was created using CUDA for the Nvidia GPUs and OpenMP for Intel MIC. DKS was further integrated into OPAL (Object-oriented Parallel Accelerator Library) in order to speed up a parallel FFT based Poisson solver and Monte Carlo simulations for particle matter interaction used for proton therapy degrader modelling. DKS was also used together with Minuit2 for parameter fitting, where χ 2 and max-log-likelihood functions were offloaded to the hardware accelerator. The concepts of the DKS, first results, and plans for the future will be shown in this paper.

Evolution of a beam dynamics model for the transport line in a proton therapy facility

“…In 2013 the Monte Carlo model for particle-matter interaction was implemented in OPAL and it will be discussed in section III B [28]. In the following 4 years, the particle-matter interaction model has been improved, benchmarked, extended with new materials and connected with a GPU (Graphics Processing Unit) card that provides a remarkable speed-up to the computation time [29]. The unique feature to combine seamless linear and nonlinear beam tracking with Monte Carlo simulations of the particle-matter interaction makes OPAL a convenient code to model a proton therapy beamline.…”

“…However, as already mentioned, the Monte Carlo computation of the particle-matter interaction is performed by means of a dedicated GPU card. This allows boosting the performance of the Monte Carlo simulation even with a high number of initial protons and a shorter time step [29].…”

On architecture and performance of adaptive mesh refinement in an electrostatics Particle-In-Cell code

“…In combination with the efficient parallelisation of such space-charge solvers using MPI (Message Passing Interface) or accelerators such as GPU (Graphics Processing Unit) and the MIC (Many Integrated Core) architecture, e.g. in [2], large-scale simulations were enabled that are more realistic. Nevertheless, multi-bunch simulations of high intensity accelerators such as cyclotrons require fine meshes in order to resolve the non-linear effects in the evolution of the beams due to space-charge.…”

Machine learning for orders of magnitude speedup in multiobjective optimization of particle accelerator systems

“…In light of these limitations, improving the execution speed and scalability of particle accelerator simulations is an area that has seen considerable effort in recent years [11,12]. Approaches to do this have focused on parallelization (e.g., see [13,14]) and hardware-based acceleration of existing simulation codes (e.g., using graphical processing units, GPUs) [15,16]. In a few exceptional cases, computationally expensive models have been used to aid live operation when on-site HPC resources are available [17].…”